Meta's Muse Wonder: Strong on agentic tasks but lags in creativity

Oct 14, 2026 · Lenny's Podcast
🎧 PodShort 24 min squeezed to 2 AI SprinklerAS AI / ML New
Episode artwork
Clara Vö
Product Lead and AI Obsessive at How I AI
Lenny's Podcast
24 min squeezed to 2
Full episode from Lenny's Podcast
Quotable Moments

AI adoption in engineering is rapidly growing, with spend on AI tools increasing 28x over the last year, but overall innovation remains flat, as new efficiencies in code generation are offset by complexities in review and validation.

Clara VöProduct Lead and AI Obsessive at How I AI
00:30

Meta's Muse Wonder is positioned as a state-of-the-art intelligence model aiming to compete in areas like personal and professional agentic work, coding, and knowledge work.

Clara VöProduct Lead and AI Obsessive at How I AI
00:08

I'm not calling it a yes, I'm just not calling it a no.

Clara VöProduct Lead and AI Obsessive at How I AI
22:58
Key Insights
  • Meta's Muse Wonder, available in two versions (Wonder and Wonderful), is positioned as a state-of-the-art intelligence model aiming to compete in areas like personal and professional agentic work, coding, and knowledge work.
  • AI adoption in engineering is rapidly growing, with spend on AI tools increasing 28x over the last year, but overall innovation remains flat, as new efficiencies in code generation are offset by complexities in review and validation.
  • Muse Wonder is particularly strong in agentic tasks and domain expertise, outperforming other models in areas like legal analysis, finance, and business workflows.
  • While Muse Wonder shows promise in coding tasks, excelling in deep-sweep and sweep-marathon benchmarks, it falls short in creative tasks like design and writing, where models like GPT-4 and Claude Opus 55 are currently superior.
  • The implementation of Muse Wonder requires subscribing to Meta's model APIs and downloading code, which can be cumbersome, though it can be run headlessly for backend tasks.
  • Muse Wonder excels at agentic work, particularly in understanding and resolving complex tasks like billing issues, managing refunds, and handling prompt injection, demonstrating a strong ability to navigate intricate workflows.
  • The model's output for emails is functional and brief, but it lacks the nuance and creativity for more complex writing tasks, making it suitable for simple, short communications.
  • While Muse Wonder performs adequately in coding and backend tasks, its creative output, particularly in design and video editing, is considered subpar compared to leading models like GPT-4 and Claude Opus 55.
Metrics Mentioned
  • Spend on AI tools grown 28x over the last year (Across more than 500 engineering organizations, as reported by DX.)

RevBots.ai View:

  • AI Sprinkler teams bolting on Muse Wonder will see backend efficiency gains.
  • ARM-stage companies need creative AI capabilities Muse Wonder lacks.
  • The 28x spend growth shows SaaS Hoarder tendencies in engineering teams.
  • Agentic strength makes it viable for ARM workflows but not creative tasks.
🎧Full Episode:Lenny's Podcast →